Meta's AIRA₃ Autonomous Research System Wins Kaggle Gold
Meta officially announced in early September that AIRA₃, its next-generation autonomous AI research system, competed in an NVIDIA-hosted Kaggle live competition in June, placing 8th among roughly 4000 teams and winning a gold medal. Meta claims this is the first Kaggle gold for an autonomous AI research system, and that the system generalizes across tasks without domain-specific redesign.
Confirmed
- The task was to fine-tune a 30B Nemotron model for stronger reasoning; all participants had identical information, ranked by held-out private-set scoring.
- AIRA₃ has no central controller: it runs many long-horizon agents (model + coding harness combos) in parallel, coordinated asynchronously via a shared forum for hypotheses/findings and a shared filesystem for artifacts.
- The 8th-place gold solution was an ensemble of GPT 5.5 (with OpenCode) and Claude 4.8 (with ClaudeCode); Meta ran further evaluations on the same private set afterwards.
- Generalization: with only the task description changed, AIRA₃ cut production GPU kernel latency by 27% on an internal benchmark, reached gold-medal level in another Kaggle competition, and translated 4,000-year-old Akkadian tablets.
Why it matters
- An autonomous research system winning gold in a fair, live competition marks a new threshold for AI-driven research capability.
- The decentralized multi-agent architecture (forum + shared filesystem) shows a viable path to organizing large-scale, long-horizon research without a central controller.
- Generalization across Kaggle, GPU kernel optimization, and ancient-language translation suggests capability beyond any single benchmark.
2026-09-06 ~ 2026-09-07 · 5 related posts
Primary sources
- Meta's AIRA₃ Coordinates Many Long-Running Agents via a Shared Forum and Filesystem — AIatMeta ·
- Meta's autonomous research agent AIRA3 wins Kaggle gold, placing 8th of ~4,000 teams — andrew_n_carr ·
- Meta's AIRA₃ Cuts GPU Kernel Latency 27% and Wins Kaggle Gold Translating 4,000-Year-Old Clay Tablets — AIatMeta ·
- GPT 5.5 + Claude 4.8 Ensemble Powered AIRA₃'s 8th-Ranked Gold Medal Kaggle Entry — AIatMeta · 2026-09-06
- [source] Meta's AIRA₃ Coordinates Many Long-Running Agents via a Shared Forum and Filesystem — AIatMeta · 2026-09-06
- [source] Meta's AIRA₃ Cuts GPU Kernel Latency 27% and Wins Kaggle Gold Translating 4,000-Year-Old Clay Tablets — AIatMeta · 2026-09-06
- [source] Meta's autonomous research agent AIRA3 wins Kaggle gold, placing 8th of ~4,000 teams — andrew_n_carr · 2026-09-06
- Meta's autonomous AI research system AIRA₃ wins Kaggle gold, ranking 8th of ~4,000 teams — TacoCohen · 2026-09-07